Pattern 4: Commit-Sized Chunks

Ship Features One Bite at a Time

Overview

A commit-sized chunk is the smallest change that still makes sense: it compiles, passes tests, and fits in one sentence.

But here’s the benefit with AI: you stay in control. AI builds one small feature, you test it, review it, understand it, then commit. The AI never gets ahead of you.

Four benefits:

  1. You control AI’s pace – one testable feature at a time, never a mysterious mess
  2. Reviews become trivial – you know exactly what each change does
  3. Deploys become safe – roll back one commit, not the whole release
  4. No blind trust needed – small diffs mean you can actually review what AI built

Teams feel the benefits first, but even solo developers stay sane when they work this way.

Example

This week I was asked to create an MCP server for the SpecStory cloud. SpecStory is a tool that captures all forms of raw intent, in this case Agent Chat history to store it. Through their cloud and their API you can leverage those chats to remember and understand what decisions were made, why, and help communicate those key decisions to others.

I created a PRD, like the one we created in chapter 2, and then I used AI to help me break that into a todo list.

:::{.callout-note title=“Todo Structure} In this case, a simple checklist sufficed. However in many cases each todo gets its own section with several pieces of information. It depends on the task :::

Key Principles

  1. One logical change – refactor OR feature, never both
  2. Always green – tests must pass
  3. Self-contained – tests, docs, and code together
  4. Narrative order – each commit continues the story

Exercise: PlantAutopsy in Commits

You have plan.md from Pattern 2. Turn it into tiny commits.

  1. Generate todos Tell AI: “From plan.md, produce todos.md—a checklist where every item is one commit-sized task.”
  2. Build first item only Ask AI to implement line 1 of todos.md—nothing else.
  3. Test and review Run it. Read the code. Understand what changed and why. If confused, ask AI to explain.
  4. Commit Only when you understand it fully: stage, test, commit.
  5. Loop Repeat. Each cycle, you’re in charge.

This way, AI never builds something you don’t understand. You review each piece while it’s small enough to grasp.

Your Turn

Slice any project into commits:

  • News scraper – Start with one site, then add deduplication and categorization in later commits
  • Tic-Tac-Toe – Build a single-file prototype, then split into game.js, ui.js, state.js
  • Spending analyzer – Load CSV in commit 1, add charts in commit 2, create summaries in commit 3

Smaller bites lead to faster ships.